JP6930195B2 - Model identification device, prediction device, monitoring system, model identification method and prediction method - Google Patents

Description

This case relates to a model identification device, a prediction device, a monitoring system, a model identification method and a prediction method.

Multiple modes may appear in the relationship between input data and output data. Therefore, a clustering technique for classifying input / output data into a plurality of modes is desired. For example, a technique for performing clustering by creating learning data for each operation mode using event information is disclosed (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2013-218725

However, in the above technique, since clustering is performed based on event information, it is difficult to perform clustering with high accuracy. As a result, it is difficult to identify the model with high accuracy.

In one aspect, the present invention aims to provide a model identification device, a prediction device, a monitoring system, a model identification method, and a prediction method capable of performing model identification with high accuracy.

In one embodiment, the model identification device uses the label addition unit that adds a label of the mode to which the model identification device belongs to a part of the plurality of input / output data, and the plurality of input / output data to which the label is attached. Of the input / output data, the clustering unit that performs clustering that classifies the input / output data with the label and the input / output data without the label into each mode, and based on the result of the clustering, of each mode A parameter estimation unit that estimates the parameters of the mathematical model and a separation superplane estimation unit that estimates the separation superplane for classifying the input data into each of the mathematical models based on the result of the clustering are provided . The clustering unit is based on the similarity between the input / output data with the label and the input / output data without the label, and the similarity between the input / output data without the label. High degree of similarity to the input / output data to which the label is attached So that the degree of similarity between the input / output data to which the label is not attached is high, and the degree of similarity to the input / output data to which the label is attached. The plurality of inputs are such that the similarity between the input / output data without the label and the input / output data without the label in the vicinity of the input / output data without the label is high. Cluster the output data.

In one embodiment, the predictor classifies the input data into each of the mathematical models using the separated superplane estimated by the separated superplane estimation unit of the model identification device, and is classified into each of the mathematical models. By applying the parameters of the mathematical model to the input data, a calculation unit for calculating the predicted value of the output data with respect to the input data is provided.

In one embodiment, the monitoring system has a determination unit that determines whether or not the degree of deviation between the predicted value calculated by the calculation unit of the prediction device and the actually measured value of the output data with respect to the input data exceeds a threshold value. Be prepared.

Model identification can be performed with high accuracy.

(A) is a diagram illustrating input / output data, and (b) is a diagram illustrating a procedure for model identification using clustering. (A) and (b) are diagrams illustrating the distribution of the relationship between the input data and the output data. It is a block diagram of the model identification apparatus which concerns on Example 1. FIG. (A) is a block diagram of the prediction device according to the first embodiment, and (b) is a block diagram of the monitoring system according to the first embodiment. It is a block diagram for demonstrating the hardware configuration of a model identification apparatus, a prediction apparatus and a monitoring system. It is a figure which illustrates the flowchart which shows the process of each part of a model identification apparatus. (A) is a diagram exemplifying index data, and (b) is a diagram exemplifying model generation data in which a mode is partially specified. (A) is a diagram exemplifying clustering using similarity, and (b) is a diagram exemplifying the result of clustering. (A) is a diagram illustrating a flowchart showing the processing of the prediction device, and (b) is a diagram illustrating a flowchart representing the processing of the monitoring system. (A) to (d) are diagrams illustrating simulation results.

First, model identification using a clustering technique will be described prior to the description of the examples. FIG. 1A is a diagram illustrating input / output data. In FIG. 1A, the horizontal axis represents the elapsed time, and the vertical axis represents the input data value and the output data value. Multiple modes may appear in the relationship between input data and output data. In the example of FIG. 1A, two types of modes, mode 1 and mode 2, appear in relation to the input data and the output data.

FIG. 1B is a diagram illustrating a procedure for model identification using clustering. As illustrated in FIG. 1B, the input / output data is classified into mode 1 and mode 2 by performing clustering of the input / output data by unsupervised learning such as k-means (step S1). Next, the mathematical model parameters of each mode are estimated using the input / output data classified into each mode (step S2). On the other hand, using the input / output data classified into each mode, the boundary surface (separation hyperplane) for classifying the unknown input data into each of the mathematical models is estimated (step S3). FIG. 1A illustrates the model parameters of the models 1 and 2 and the boundary surface. According to this procedure, unknown input data can be classified into each mathematical model. In addition, the model parameters of each mathematical model can be used to predict the output data for unknown input data.

In this procedure, the result of clustering in step S1 has a great influence on the prediction accuracy of the output data. Therefore, it is desired to improve the accuracy of clustering. However, the result of clustering may vary depending on the distribution of input / output data. The reason will be described below.

2 (a) and 2 (b) are diagrams illustrating the distribution of the relationship between the input data and the output data. In FIGS. 2A and 2B, u (t-1) is input data to the measurement target of the sensor, and y (t) is output data of the sensor. In the examples of FIGS. 2 (a) and 2 (b), the output data value increases linearly as the input data value increases, increases linearly as the input data value increases from the middle, and then decreases linearly, and then inputs from the middle. As the data value increases, it increases linearly. Depending on how it is perceived, an input / output relationship that is interrupted in the middle but increases linearly can be classified into mode 1, and an input / output relationship that decreases linearly can be classified into mode 2. FIG. 2A shows the result of classifying the input / output relationship into mode 1 and mode according to this classification. Depending on another way of thinking, an input / output relationship that increases linearly and decreases linearly from the middle can be classified into mode 1, and an input / output relationship that increases linearly can be classified into mode 2. FIG. 2B shows the results of classifying the input / output relationship into mode 1 and mode 2 according to this classification.

The clustering results themselves are the cases of FIGS. 2 (a) and 2 (b), both of which can be said to be correct results. However, there may be cases where the result does not match the actual mode of the I / O relationship. In this case, the accuracy of model identification may decrease, and the accuracy of prediction of output data may decrease. Therefore, it is desired to improve the accuracy of model identification by increasing the accuracy of clustering.

Therefore, in the following examples, a model identification device, a prediction device, a monitoring system, a model identification method, and a prediction method capable of performing model identification with high accuracy will be described.

FIG. 3 is a block diagram of the model identification device 100 according to the first embodiment. FIG. 4A is a block diagram of the prediction device 200 according to the first embodiment. FIG. 4B is a block diagram of the monitoring system 300 according to the first embodiment.

As illustrated in FIG. 3, the model identification device 100 includes a sensor data acquisition unit 11, a sensor data storage unit 12, a model generation data acquisition unit 13, a label data integration unit 14, a clustering calculation unit 15, and a model parameter estimation unit 16. , A separation superplane estimation unit 17, a model storage unit 18, and a label addition unit 20. The label adding unit 20 includes an index data acquisition unit 21, an index data display unit 22, a label information acquisition unit 23, and a label data creation unit 24.

As illustrated in FIG. 4A, the prediction device 200 includes a sensor data acquisition unit 11, a model storage unit 18, a model acquisition unit 31, a prediction input data acquisition unit 32, a prediction calculation unit 33, and a prediction value storage unit. 34 is provided. As illustrated in FIG. 4B, the monitoring system 300 includes a sensor data acquisition unit 11, a predicted value storage unit 34, and a determination unit 40.

FIG. 5 is a block diagram for explaining the hardware configurations of the model identification device 100, the prediction device 200, and the monitoring system 300. As illustrated in FIG. 5, the model identification device 100, the prediction device 200, and the monitoring system 300 include a CPU 101, a RAM 102, a storage device 103, an input device 104, a display device 105, and the like. Each of these devices is connected by a bus or the like. The CPU (Central Processing Unit) 101 is a central processing unit. The CPU 101 includes one or more cores. The RAM (Random Access Memory) 102 is a volatile memory that temporarily stores a program executed by the CPU 101, data processed by the CPU 101, and the like. The storage device 103 is a non-volatile storage device. As the storage device 103, for example, a ROM (Read Only Memory), a solid state drive (SSD) such as a flash memory, a hard disk driven by a hard disk drive, or the like can be used. The input device 104 is a keyboard, a mouse, or the like. The display device 105 is a liquid crystal display, an electroluminescence panel, or the like, and displays the processing results of the model identification device 100 and the prediction device 200. The model identification device 100, the prediction device 200, and the monitoring system 300 are realized by the CPU 101 executing the program stored in the storage device 103. The model identification device 100, the prediction device 200, and the monitoring system 300 may be hardware such as a dedicated circuit.

FIG. 6 is a diagram illustrating a flowchart showing processing of each part of the model identification device 100. Hereinafter, the processing of the model identification device 100 will be described with reference to FIGS. 3 and 6.

First, the sensor data acquisition unit 11 acquires the input data and the output data of the sensor with respect to the input data as sensor data (step S11). The sensor data acquired by the sensor data acquisition unit 11 is stored in the sensor data storage unit 12.

Next, the index data acquisition unit 21 acquires index data for model generation from the sensor data stored in the sensor data storage unit 12. FIG. 7A is a diagram illustrating index data. As illustrated in FIG. 7A, the index data has not yet been classified into each mode because it is before clustering. The index data display unit 22 causes the display device 105 to display the index data (step S12).

Next, the user who confirms the content displayed on the display device 105 inputs the label information for some index data to the input device 104. The label information acquisition unit 23 acquires label information from the input device 104. The label data creation unit 24 creates label data based on the label information acquired by the label information acquisition unit 23 (step S13). The label data is, for example, a mode identifier for each of the index data. For example, for each sensor data, "1" representing mode 1, "2" representing mode 2, and "0" not specifying a mode are created.

The label information may be entered by the user or may be entered automatically. For example, by predetermining the rough relationship between the input data and the output data, the label of the index data can be automatically specified. For example, in the example of FIG. 7A, a mathematical formula assuming that the model is switched at a point where the alignment is interrupted is determined in advance. In the example of FIG. 7A, for example,
y (t) = u (t-1) + 2 + e (t), -4 ≦ u (t-1) ≦ -1
y (t) = -u (t-1) + e (t), -1 <u (t-1) <2
y (t) = u (t-1) + 2 + e (t), 2 ≦ u (t-1) ≦ 4
Three mathematical formulas can be defined. Note that e (t) is white noise. Label information for some index data can be automatically created by using the relationship of this mathematical formula.

The model generation data acquisition unit 13 acquires model generation data from the sensor data stored in the sensor data storage unit 12. For example, the model generation data is the same as the index data acquired by the index data acquisition unit 21. Next, the label data integration unit 14 integrates the model generation data and the label data. As a result, the mode is specified as part of the model generation data. As illustrated in FIG. 7B, the mode is specified for some model generation data. In the example of FIG. 7B, the diamond represents mode 1 and the cross represents mode 2.

Next, the clustering calculation unit 15 sets parameters for performing clustering (step S14). Next, the clustering calculation unit 15 performs a clustering calculation using the model generation data received from the label data integration unit (step S15).

The clustering calculation unit 15 performs clustering using the labeled data. For example, the clustering calculation unit 15 performs clustering based on the similarity between the labeled data and the unlabeled data and the similarity between the unlabeled data. Specifically, as illustrated in FIG. 8A, the clustering calculation unit 15 increases the similarity between the unlabeled data, which has a high degree of similarity to the labeled data, in the model generation data. In addition, clustering is performed so that the similarity between the unlabeled data, which has a low degree of similarity to the labeled data, and the data in the vicinity thereof are high.

For example, the clustering calculation unit 15 uses the learning standard (evaluation function) of the following equation (1) for clustering. In the following equation (1) _{, Wi and i'are} similarity functions, and are functions that take a large value when the distance between them is short. In the parentheses of the following equation (1), the first term represents the learning of the discriminative model using labeled data (unsupervised learning), the second term is the normalization term that suppresses overlearning, and the third term is the label. Represents learning of a discriminative model using no data (unsupervised learning). The discriminative model can be calculated by the following formula (1). The discriminative model can be expressed as the following equation (2). In the following equation (2), K (x, x _j ) is a feature quantity (kernel) of the input x. By obtaining the discriminative model, the discriminative formula of the following formula (3) can be obtained. The sensor data can be classified into each mode by discriminating between -1 and +1 using the identification formula. FIG. 8B shows the results of clustering using the following equations (1) to (3).

The clustering calculation unit 15 classifies each of the model generation data into each mode by using the above equations (1) to (3) (step S16). The above processing completes the data classification flowchart.

Next, the model parameter estimation unit 16 determines the input / output data for model creation (step S21). Specifically, the model parameter estimation unit 16 determines the model generation data for each mode obtained in step 16 as the model creation input / output data. Next, the model parameter estimation unit 16 determines a mathematical model structure suitable for model creation input / output data from a plurality of types of mathematical model structures created in advance (step S22). Next, the model parameter estimation unit 16 determines the order of the mathematical model (step S23). The degree of a model is the degree of a mathematical formula in a mathematical model. Next, the model parameter estimation unit 16 estimates the model parameters using the model generation data for each mode obtained in step S16 (step S24).

On the other hand, the separation hyperplane estimation unit 17 determines a classifier suitable for the model generation data for each mode obtained in step S16 from a plurality of types of classifiers (step S31). Next, the separation hyperplane estimation unit 17 determines whether or not the classifier determined in step S31 can separate the model generation data into each mode (step S32). If "No" is determined in step S32, the process is executed again from step S31, and another classifier is determined. When it is determined as "Yes" in step S32, the separation hyperplane estimation unit 17 sets the classifier parameter (step S33). Next, the separation hyperplane estimation unit 17 generates a classifier (step S34). A classifier is a hyperplane separation. The model parameters estimated in step S24 and the classifier generated in step S34 are stored in the model storage unit 18.

Next, the details of the processing of the prediction device 200 will be described with reference to FIGS. 4 (a) and 9 (a). FIG. 9A is a diagram illustrating a flowchart showing the processing of the prediction device 200. The model acquisition unit 31 acquires the classifier and model parameters stored in the model storage unit 18 (step S41). Next, the prediction input data acquisition unit 32 acquires the prediction input data from the sensor data acquisition unit 11 (step S42). The input data for prediction is the input data after the model generation data.

Next, the prediction calculation unit 33 classifies the input data for prediction into each model by using the classifier (step S43). Next, the prediction calculation unit 33 calculates the prediction value of the output data by applying the model parameters to the input data for prediction classified into each model (step S44). The prediction calculation unit 33 stores the calculated predicted value in the prediction value storage unit 34 (step S45).

Next, the details of the processing of the monitoring system 300 will be described with reference to FIGS. 4 (a) and 9 (b). FIG. 9B is a diagram illustrating a flowchart showing the processing of the monitoring system 300. The determination unit 40 acquires the output data (measured value) corresponding to the input data acquired in step S42 from the sensor data acquisition unit 11, and acquires the predicted value stored in the predicted value storage unit 34 (step S51). ).

Next, the determination unit 40 determines whether or not the degree of deviation between the measured value and the predicted value is equal to or greater than the threshold value (step S52). For example, a statistical value such as 3σ can be used as the threshold value. If "No" is determined in step S52, the determination unit 40 outputs information related to normality (step S53). If it is determined as "Yes" in step S52, the determination unit 40 outputs information related to the abnormality (step S54).

According to this embodiment, a label of the mode to which the input / output data belongs is added to a part of the plurality of input / output data. Since the clustering is performed to classify the plurality of input / output data into each mode using the input / output data to which the label is attached, the input / output data to which the label is attached can be used as prior knowledge. As a result, the clustering accuracy is improved as compared with the case where prior knowledge is not used. As a result, the model identification accuracy can be improved. In addition, the higher model identification accuracy improves the estimation accuracy of the separated hyperplane and model parameters. As a result, the prediction accuracy of the predicted value of the output data with respect to the unknown input data is improved. As a result, the prediction accuracy of abnormality judgment based on the degree of deviation between the predicted value and the measured value is improved.

10 (a) to 10 (d) are diagrams illustrating simulation results. FIG. 10A shows the result when clustering was performed without adding a label to the sensor data. As illustrated in FIG. 10A, when no label is added to the sensor data, the input / output data group that increases linearly and decreases linearly from the middle is classified into mode 1, and is further linearly increased. The input / output data group is classified into mode 2. FIG. 10B shows the results when clustering is performed by adding labels to some sensor data. As illustrated in FIG. 10B, the input / output data group that is interrupted in the middle but becomes linearly large is classified into mode 1, and the input / output data group that becomes linearly small is classified into mode 2. The result of FIG. 10B is the result of correct clustering.

A classifier was created based on the results of this clustering, and model parameters were estimated. Then, the unknown input data was classified into each model using a classifier, and the output data for the input data was predicted using the model parameters of each model. FIG. 10 (c) shows a predicted value and an actually measured value when the result of FIG. 10 (a) is used. FIG. 10 (d) shows a predicted value and an actually measured value when the result of FIG. 10 (b) is used. RMSE (Root Mean Square Error) was calculated as the error between the predicted value and the measured value. In the example of FIG. 10 (c), the RMSE was 0.64. In the example of FIG. 10D, the RMSE was 0.13. When the output data was predicted 10 times in total for different unknown input data, the RMSE averaged 6.51 when the sensor data was not labeled. When a label was added to the sensor data, the RMSE averaged 0.60. As described above, when the label was added to the sensor data, the prediction accuracy was improved by 90.1%.

In the above example, for example, the input data is an input value to a device of a predetermined facility, and the output data is an output value of a sensor provided in the facility. Specifically, the input data is an input value for the temperature control device in a predetermined plant, and the output data is an output value of the temperature sensor of the plant. However, the input data and the output data are not particularly limited as long as they have a causal relationship with each other. For example, the input data may be the output value of a predetermined sensor, and the output data may be the output value of another sensor having a causal relationship with the output value of the sensor.

In each of the above examples, the label addition unit 20 functions as an example of the label addition unit that adds the label of the mode to which the label belongs to a part of the plurality of input / output data. The clustering calculation unit 15 functions as an example of a clustering unit that performs clustering that classifies the plurality of input / output data into each mode using the input / output data to which the label is attached. The model parameter estimation unit 16 functions as an example of a parameter estimation unit that estimates the parameters of the mathematical model of each mode based on the result of the clustering. The separation hyperplane estimation unit 17 functions as an example of the separation hyperplane estimation unit that estimates the separation hyperplane for classifying the input data into each of the mathematical models based on the result of the clustering. The prediction calculation unit 33 classifies the input data into each of the mathematical models using the separated superplane estimated by the separated superplane estimation unit, and the mathematical model is added to the input data classified into each of the mathematical models. By applying the parameters of, it functions as an example of a calculation unit that calculates a predicted value of output data with respect to the input data. The determination unit 40 functions as an example of a determination unit that determines whether or not the degree of deviation between the predicted value calculated by the calculation unit and the actually measured value of the output data with respect to the input data exceeds the threshold value.

Although the examples of the present invention have been described in detail above, the present invention is not limited to the specific examples, and various modifications and modifications are made within the scope of the gist of the present invention described in the claims. It can be changed.

11 Sensor data acquisition unit 12 Sensor data storage unit 13 Model generation data acquisition unit 14 Label data integration unit 15 Clustering calculation unit 16 Model parameter estimation unit 17 Separated superplane estimation unit 18 Model storage unit 20 Label addition unit 21 Index data acquisition unit 22 Index data display unit 23 Label information acquisition unit 24 Label data creation unit 31 Model acquisition unit 32 Prediction input data acquisition unit 33 Prediction calculation unit 34 Prediction value storage unit 40 Judgment unit 100 Model identification device 200 Prediction device 300 Monitoring system

Claims (5)

A label addition part that adds a label of the mode to which it belongs to a part of multiple input / output data,
Using the input / output data with the label attached, clustering is performed to classify the input / output data with the label and the input / output data without the label among the plurality of input / output data into each mode. The clustering part to be performed and
A parameter estimation unit that estimates the parameters of the mathematical model of each mode based on the result of the clustering,
A separated hyperplane estimation unit for estimating a separated hyperplane for classifying input data into each of the mathematical models based on the result of the clustering is provided .
The clustering unit is based on the similarity between the labeled input / output data and the unlabeled input / output data, and the similarity between the unlabeled input / output data. , High degree of similarity to the input / output data to which the label is attached So that the degree of similarity between the input / output data to which the label is not attached is high, and similarity to the input / output data to which the label is attached. The plurality of input / output data having a low degree of similarity between the unlabeled input / output data and the unlabeled input / output data in the vicinity of the unlabeled input / output data are increased. A model identification device characterized by clustering input / output data. The input data is classified into each of the mathematical models using the separated superplane estimated by the separated superplane estimation unit of the model identification device according to claim 1, and the input data classified into each of the mathematical models. A prediction device comprising a calculation unit that calculates a predicted value of output data with respect to the input data by applying the parameters of the mathematical model to the above. A determination unit for determining whether or not the degree of deviation between the predicted value calculated by the calculation unit of the prediction device according to claim 2 and the actually measured value of the output data with respect to the input data exceeds a threshold value is provided. Monitoring system. To some of the plurality of input and output data, it appends the label belongs to mode,
Using the input / output data with the label attached, clustering is performed to classify the input / output data with the label and the input / output data without the label among the plurality of input / output data into each mode. There line,
Based on the results of the clustering to estimate the parameters of the mathematical model of each mode,
Based on the results of the clustering to estimate the separating hyperplane for classifying each of said mathematical model input data, process the computer has performed,
When performing the clustering, the similarity between the input / output data to which the label is attached and the input / output data to which the label is not attached, and the similarity between the input / output data to which the label is not attached are determined. Based on this, the degree of similarity with the input / output data to which the label is attached is high so that the degree of similarity between the input / output data without the label is high, and the degree of similarity with the input / output data to which the label is attached is high. The similarity between the unlabeled input / output data and the unlabeled input / output data in the vicinity of the unlabeled input / output data is high. A model identification method characterized by clustering a plurality of input / output data. The input data is classified into each of the mathematical models using the separated hyperplane estimated by the model identification method according to claim 4, and the parameters of the mathematical model are added to the input data classified into each of the mathematical models. A prediction method, characterized in that a computer executes a process of calculating a predicted value of output data with respect to the input data by applying.

Images